Stoch Environ Res Risk Assess
DOI 10.1007/s00477-013-0821-z
REVIEW PAPER
An interactive biplot implementation in R for modeling
genotype-by-environment interaction
Elisa Frutos • M. Purificación Galindo •
Vı́ctor Leiva
 Springer-Verlag Berlin Heidelberg 2013
Abstract Classical and GGE biplot methods are graphi-
cal procedures that allow multivariate data to be analyzed.
In particular, the GGE biplot displays the genotype main
effect (G) and the genotype by environment interaction
(GE) in two-way data. The GGE biplot originates from data
graphical analysis of multi-environment trials (MET).
Thus, agronomists, crop scientists and geneticists are
potential users of this method. However, it can also be used
to visualize and analyze other types of data. In this paper,
we propose a new interactive computational implementa-
tion in R language to perform the main functions of the
classical and GGE biplot methods, so it is also useful for
MET data visual analysis. This implementation is orga-
nized in an R package named GGEBiplotGUI. This
package is the only interactive, noncommercial and open
source software that currently exists, offering a free alter-
native to available commercial software. In addition, it can
be used without to practically have knowledge of the R
programming language. Here, we present and discuss the
capabilities and features of the GGEBiplotGUI package
and illustrate them by using real data. The GGEBiplot-
GUI package graphically addresses the questions that a
researcher likely asks. This R package is not only a tool for
visual data analysis of multi-environment trials, useful for
plant breeders and geneticists, in order to study yields from
E. Frutos  M. P. Galindo
Departamento de Estadı́stica, Universidad de Salamanca,
Salamanca, Spain
V. Leiva (&)
Departamento de Estadı́stica, Universidad de Valparaı́so, Avda.
Gran Bretaña 1111, Playa Ancha, Valparaiso, Chile
e-mail: victor.leiva@uv.cl
URL: http://www.deuv.cl/leiva
genotypes and interactions between genotype and envi-
ronment, but also data from other areas can be analyzed by
the GGEBiplotGUI package.
Keywords Graphical display  Least squares
method  Multivariate data analysis  Singular value
decomposition  Statistical models  Statistical
software
Mathematics Subject Classification 62H99 
62-04
1 Introduction and bibliographical review
The biplot method is a graphical display of multivariate
data. This method was introduced by Gabriel (1971) in the
context of principal component analysis (PCA). Specifi-
cally, the biplot is a joint graphical representation, in a low
dimensional Euclidean space (usually a plane), of a mul-
tivariate data matrix by markers for its rows and columns,
chosen in such a way that the inner (or scalar) product
represents the elements of the data matrix. Due to the
properties of the inner product, the biplot is a powerful data
visualization tool, which can be viewed as a multivariate
extension of the scatterplot, because the biplot represen-
tation is usually performed in the two-dimensional space of
the set of real numbers. This is the classical focus of the
biplot method (classical biplot), which has two parts. First,
it carries out the approximation of the data matrix by a
singular value decomposition (SVD) and, second, this
matrix is factorized in row and column markers; see Eckart
and Young (1939) and Golub and Reinsch (1970).
Gower and Harding (1988), Gower (1992), and Gower
and Hand (1996) provided a different focus of the biplot to
123

that originally proposed by Gabriel (1971). They first
ordered the individuals using scaling, and then superposed
the variables in a such a way that a joint graphical inter-
pretation is possible, as usual in biplot methods. Both of
these representations are just descriptive, so that no para-
metrical assumptions are considered. For more details
about the biplot methods, see Gower and Hand (1996),
which is recognized as the first book on this method, and
Greenacre (2010) and Gower et al. (2011), for more recent
books.
Another focus different from the classical one is based
on the relation between the approximation of least squares
(LS) of matrices and the SVD. This third focus permits the
biplots to be visualized as models fitted by bilinearity.
These bilinear models are known as regression biplots and
can be interpreted as a multiplicative bilinear model,
allowing us to demonstrate in graphical form the associa-
tion, linear or not, between subjects and/or variables; see
Gabriel (1998) and Vicente-Villardón et al. (2006). In this
line, a number of authors have studied the case when the
biplots are used to describe interaction between row and
column effects in multiplicative and additive bilinear
models; see, e.g., Denis (1991), Falguerolles (1995), Van
Eeuwijk (1995), and Choulakian (1966). In order to
approximate the biplots in the case of these last models, the
mentioned authors used estimation methods employed in
generalized bilinear models, which are considered as an
extension of the generalized linear models; see Nelder and
Wedderburn (1972). An exhaustive study on all of these
alternative ways of the biplot method can be seen in
Cárdenas and Galindo (2003).
Since Gabriel (1971)’s paper, the biplot method has
increasingly been used in data visualization and analysis in
diverse disciplines, being it particularly useful in exploring
data from areas such as agricultural, ecological and environ-
mental sciences, and genetics. In these areas, the size of the
data sets often result to be large, with complicated interactions
and interconnections. These sets are typically genotype-by-
environment-by-trait three-way data, which provide rich
information. Such data can be organized into several types of
two-way tables, as for example: (i) genotype 9 environment
tables for each trait; (ii) genotype 9 trait tables for each
environment, across subsets of environments, and across all
environments; (iii) environment 9 trait tables for each
genotype, across subsets of genotypes, and across all geno-
types; (iv) phenotype (genotype–environment combina-
tion) 9 trait tables; (v) genotype 9 environment-trait tables;
and (vi) combined two-way tables of genotype 9 explanatory
variables plus response variables. A full understanding of the
type of data requires knowledge of all these two-way tables.
From the viewpoint of genotype evaluation, genotype by
environment and genotype by trait tables are the most relevant;
see Yan et al. (2000).
123
Stoch Environ Res Risk Assess
We focus our attention on data of genotype by environ-
ment, which provide rich information for different purposes,
but we want to stress the methodology discussed in the paper
can also be applied to several other types of data, for
example in geosciences; see Bacon-Shone (2009). Plant
geneticists often study the performance of many genotypes
in diverse environments. Such studies are conducted in order
to select the best genotypes for improvements of crops. The
data collected for these studies correspond to one or more
attributes, for each genotype in each environment. This type
of data can be analyzed in a straightforward way by means
of analysis of variance (ANOVA) models. However, other
models can also be used. Particularly, the additive main
effects and multiplicative interaction (AMMI) model and
genotype main effects (G) and genotype 9 environment
interaction effects (GE) model, known as GGE, are the most
used to describe the kind of data above mentioned; see
Gauch (1988). The AMMI model (AMMI biplot) has been
extensively applied in the statistical analysis of multiple-
environment trials (MET); see Kempton (1984), Gauch and
Zobel (1996, 1997), and Ebdon and Gauch (2002a, b). The
GGE model (GGE biplot) was proposed by Yan et al.
(2000), allowing genotype by environment interaction (GEI)
of MET data to be visually examined. Several recent papers
have exhaustively compared and contrasted AMMI and
GGE models, with respect to their suitability for genotype
by environment interaction analysis; see Gauch (2006), Yan
and Tinker (2006), Yan et al. (2007) and Gauch et al.
(2008). In AMMI models, as multiplicative terms as needed
are incorporated in order to explain the variability of the
second order interaction. Such models are based on the
decomposition in singular values and vectors of the residual
matrix of the associated linear model. The GGE model
applies the SVD to the data, subtracting the environment
effects, because these biplots display both G and GE, which
are the two sources of variation relevant to cultivar evalu-
ation; see Kang (1988), Gauch and Zobel (1996) and Yan
and Kang (2003). This model is based on multiplicative
linear-bilinear site regression; see Cornelius et al. (1996).
Bibliography on biplots is generous, with many papers
published in various scientific journals, whereas about
50,000 websites containing the word biplot are available in
the internet. Macros for biplot analysis have been imple-
mented in the main commercial statistical softwares; see,
e.g., Yan and Tinker (2006). Currently, most commercial
statistical software packages include a procedure or macro
for generating biplots. Specifically, the GGEbiplot pro-
gram, dedicated to the GGE biplot (see http://www.
ggebiplot.com), can also generate the classical biplot.
This program is a commercial software and is widely used
by agronomists, crop scientists and geneticists; see Yan and
Kang (2003, 2006). However, today the scientific com-
munity has at its disposal a non-commercial and open
Stoch Environ Res Risk Assess
source software for statistics and graphs, named R, which
can be obtained at no cost from http://www.r-project.org.
The statistical software R is currently very popular in the
international scientific community; see R Team (2013), and
Leiva et al. (2008) and Barros et al. (2009), for some R
packages that have been developed. Thus, it seems to be
necessary to count with an R package that offers a non-
commercial and open source software as alternative to the
existing commercial software on biplots. This R package
should address the main questions that a researcher could
ask and have an interactive characteristic, because in this
way it can be more attractive for practitioners that does not
know the R programming language.
The aims of this work are (i) to propose a new inter-
active computational implementation in R language to
perform the main functions of the GGE and classical biplot
methods; and (ii) to illustrate it by using real data related to
genotype-by-environment interaction. This implementation
is incorporated by an R package named GGEBiplotGUI,
which is a free alternative to the commercial software
created by Yan and Tinker (2006) and that can be down-
loaded from http://cran.r-project.org. The GGEBiplot-
GUI package is a graphical interface that allows us to use it
without necessity of having a knowledge of the R language,
which can do this package very attractive for diverse
practitioners. The different functions of this package are
organized by means of menu options that allows them to be
used in a friendly and easy way. Although the main
objective of the GGEBiplotGUI package is to construct
and manipulate GGE biplots, it is also possible to construct
by this package AMMI biplots, only selecting the correct
model, as well as the classical biplot. In addition, the
GGEBiplotGUI package can be used to analyze geno-
type-by-environment data because it includes: (i) mega-
environment analysis; (ii) test-environment evaluation; and
(iii) genotype evaluation. In order to illustrate the capa-
bilities and features of the GGEBiplotGUI package, we
use the proposed implementation to analyze real agricul-
tural and genomic data.
The paper is organized as follows. In Sect. 2, we provide
the theoretical background of this work. In Sect. 3, we
propose and discuss the new computational implementa-
tion in R language for biplot methods. In Sect. 4, we per-
form the empirical application of the proposed
computational implementation by using real data. Finally,
in Sect. 5, we provide some conclusions.
2 Background
In this section, we provide an overview on classical and
GGE biplot methods.
2.1 Classical biplots
Any n 9 m matrix Y ¼ ðyij Þ of rank r can be factorized
into an n 9 r matrix G and an m 9 r matrix H; both
necessarily of rank r, that is, as
Y ¼ GH> : ð1Þ
Factorization given in Eq. (1) can be written in a scalar way as
yij ¼ g> ð2Þ
i hj ;
for each i and j, where yij is the element in the ith row and
jth column of Y; g> i is the ith row of G and hj is the jth row
of H: Factorization given in Eq. (2) assigns each vector
g1 ; . . .; gn (row effects) to each one of the n rows of Y; and
each vector h1 ; . . .; hm (column effects) to each column of
Y: Thus, Eq. (2) provides a representation of Y by means
n ? m vectors in the r-dimensional space. For a matrix of
rank equal to two, these n ? m vectors may be plotted in
the plane, giving a representation of the n m elements of Y;
by means of the inner product of the corresponding row
effect and column effect vectors. Such a plot is referred to
as a biplot, since it allows row effects and column effects to
be jointly plotted; for more details, see Gabriel (1971).
Matrices with ranks greater than two cannot be repre-
sented exactly by a biplot. However, if a matrix Y can be
approximated by a matrix of rank equal to two, say Y ð2Þ ;
the inner products of the row and column effects can
approximate the elements of Y:
To approximate any n 9 m matrix Y of rank r by an
n 9 m matrix of smaller rank, one may use the SVD,
which can, in this case, be written as
X r
Y ¼ UKV > ¼ kk u k v >
k; ð3Þ
k¼1
where U ¼ ðu1 ;...;ur Þ;V ¼ ðv1 ;...;vr Þ and K ¼ ðk1 ;...;kr Þ;
with vectors uk and v>k being called left-singular and right-
singular vectors, respectively, and kk being the singular
values. For each k ¼ 1;...;r; the singular value kk and the
singular vectors uk and vk are chosen to satisfy
u>
k Y ¼ kk v k
Yvk ¼ kk uk
YY > uk ¼ k2k uk
Y > Yvk ¼ k2k vk ;
with k1  k2      kr ; such that u> >
k ul ¼ vk vl ¼ dkl ;
where dkl denotes the Kronecker delta. Note that, by using
the LS method, expression given in Eq. (3) can be written as
X
s
Y ðsÞnm ¼ U ðsÞns KðsÞss V >
ðsÞsm ¼ kk uk v>
k; ð4Þ
k¼1
which corresponds to the best approximation for Y of rank
equal to s (see Householder and Young 1938), that is,
123

expression in Eq. (4) provides an n 9 m matrix, say
M ¼ ðmij Þ; of rank s that minimizes k Y  M k2 ¼
Pn Pm 2 2 2 2
i¼1 j¼1 ðyij  mij Þ : Because k Y k ¼ k1 þ    þ kr ; the
goodness of fit of the model to the data can be measured by
P P
( sk=1k2k / rk=1k2k ) 100 %.
It is also possible to obtain the marker matrices G and H
by means of alternating regressions. Specifically, if we
consider the row markers G as fixed, the column markers
can be computed by regression as (i) H> ¼ ðG> GÞ1 G> Y:
In the same way, fixing H; the row markers can be obtained
as (ii) G> ¼ ðH> HÞ1 H> Y: Thus, alternating steps (i) and
(ii), their products converge to the SVD; for details about
alternating regressions, see Ukkelberg and Borgen (1993).
By choosing factors G and H of Y ð2Þ ; as in Eq. (1), for
biplotting, one may use the factorization provided by the
SVD given by G ¼ UKc and H ¼ VK1c ; where the scalar
c can take any value between zero and one. When c = 1,
the singular values are entirely partitioned into the row
eigenvectors, which is referred to as row-metric preserving
or JK biplot; see Gabriel (1971). Because in this case G ¼
UK; we have YY > ¼ GG> : Therefore, this partitioning
recovers the Euclidean distances among row factors and is,
consequently, appropriate for visualizing the similarity/
dissimilarity among row factors. When c = 0, the singular
values are entirely partitioned into the column eigenvec-
tors, referred to as column-metric preserving or GH biplot;
see Gabriel (1971). Because now H ¼ VK; we have
Y > Y ¼ HH> ; which is the sum of squares and cross pro-
ducts matrix of Y: If Y is column-centered, then Y > Y is
(m - 1) times the covariance matrix. Thus, this partition-
ing is appropriate for studying the relationships among
column factors. Two important rules applying to such a
partitioning are (i) the correlation between two columns is
approximated by the cosine of the angle between their
vectors, if the data are column-centered before subjecting
to SVD; and (ii) the vector length of a column equals
pffiffiffiffiffiffiffiffiffiffiffiffi
m  1 times the standard deviation (SD) of the column
factor, across the rows. When c = 0.5, symmetric roles are
assigned to rows and columns, referred to as SQRT biplot
or symmetric biplot; see Cárdenas and Galindo (2003).
2.2 The HJ biplot
Galindo (1986) proposed a new form of representation,
known as HJ biplot, in which the coordinates for columns
coincide with the column markers in the GH biplot,
whereas the coordinates for the rows coincide with the row
markers in the JK biplot. Note that these coordinates may
be represented in the same Cartesian system.
The HJ biplot is conceptually similar to the correspon-
dence analysis, but it applies to continuous data rather than
123
Stoch Environ Res Risk Assess
to categorical data. The rules for the interpretation of the HJ
biplot are a combination of the rules used in classical bi-
plots, correspondence analysis, factor analysis and multi-
dimensional scaling techniques. Specifically, we have that:
(i) the distances among row markers are interpreted as
an inverse function of similarities, in a such a way
that closer markers (individuals) are more similar;
this property allows the clusters of individuals with
similar profiles to be identified;
(ii) the lengths of the column markers (vectors) approx-
imate the SD of the variables;
(iii) the cosines of the angles among the column vectors
approximate the correlations among variables in a
such a way that small acute angles are associated
with variables that have high positive correlations,
obtuse angles near to the straight angle are associated
with variables that have high negative correlations,
and right angles are associated with non-correlated
variables; in the same way, the cosines of the angles
among the variable markers and the axes (principal
components) approximate the correlations between
them; whereas for standardized data, they approxi-
mate the factor loadings in factor analysis; and
(iv) the order of the orthogonal projections of the row
markers (points) onto a column marker (vector)
approximates the order of the row elements (values)
in that column; so, as the projection of a point
(individual) away from the center of gravity (average
coordinate point), the value that this individual takes
on the variable is farther from its mean.
2.3 The GGE biplot
GEI is commonly observed by crop producers and breeders
as the differential ranking of cultivar yields among loca-
tions or years; see Samonte et al. (2005). Plant breeders
conduct MET primarily to identify the superior cultivar for
a target region and secondarily to determine whether the
target region can be subdivided into different mega-envi-
ronments; see Yan et al. (2000). Because the main objec-
tive of a breeding program is to select genotypes that are
consistent and highly productive across different environ-
ments, the existence of GEI presents a problem for
breeders.
In practice, standard statistical methods that have been
applied for analysis of GEI include:
(i) ANOVA model;
(ii) completely multiplicative model (COMM);
(iii) shifted multiplicative model (SHMM);
(iv) genotypes (cultivars) regression model (GREG);
(v) sites (environments) regression model (SREG); and
Stoch Environ Res Risk Assess

(vi) AMMI model. Cornelius et al. (1992) and Crossa and Cornelius (1997).
On the one hand, ANOVA is an additive model that The GREG is a reparameterization of the stability analysis
describes the main effects and determines whether GEI is a model of Finlay and Wilkinson (1963) and Eberhart and
significant source of variation or not. However, it does not Russell (1969). This model expresses the performance of
provide insight into the genotypes or environments that each genotype as its multiple regression on t unknown
give rise to the interaction; see details in Samonte et al. functions gjk of unidentified characteristics of the sites; see
(2005). Specifically, for the data matrix Y ¼ ðyij Þ; with Cornelius et al. (1996). In this context, we use the name
response variables Yij, the ANOVA model is ‘‘genotype regression’’ instead of ‘‘row regression’’, men-
tioned in Bradu and Gabriel (1978). Similarly, we use the
Yij ¼ l þ ai þ bj þ /ij þ eij ; ð5Þ name ‘‘site regression’’ instead of ‘‘column regression’’.
where Yij is the yield of the genotype i in the environment The multiplicative terms in the models given in
j, l is the overall mean, ai is the genotype (row) main Eqs. (6)–(10) are estimated by the LS method through the
effect, bj is the environment (column) main effect, /ij is the SVD of the two-way array of data (matrix Z), obtained by
specific genotype i (row) by the environment j (column) subtracting the LS estimates from the additive effects of the
interaction, and eij is the error term of the model, where model, using the cell means; see Cornelius et al. (1996).
iid For these five models, the LS estimates of the additive
eij  Nð0; r2 Þ: On the other hand, COMM, SHMM, GREG, effects and expressions for the elements of the appropriate
SREG and AMMI are multiplicative model forms; see Z ¼ ðzij Þ matrix are: [COMM] zij ¼ yij ; [SHMM] zij ¼
Cornelius et al. (1996). Specifically, the COMM is
yij  b l ¼ y:: ; [GREG] zij ¼ yij  b
l ; with b a i ; with b
a i ¼ yi: ;
X
t
Yij ¼ kk nik gjk þ eij ; ð6Þ [SREG] zij ¼ y  b
ij b ; with b
j b ¼ y ; and [AMMI] zij ¼
j :j
k¼1 ai  b
yij  b l ; with b
bj  b l; b
a i ¼ yi:  b b j ¼ y:j  b
l ; and b
l¼
the SHMM is y:: The AMMI model first applies the ANOVA model and
then applies the PCA model to the interaction; see Gauch
X
t
(1988). Then, the genotypic and environmental scores can
Yij ¼ l þ kk nik gjk þ eij ; ð7Þ
k¼1 be used to construct biplots that help to interpret the GEI.
GGE applies SVD to the environment-centered two-way
the GREG is data containing G and GE, but the AMMI model applies
X
t the SVD to the doubly-centered two-way data, containing
Yij ¼ ai þ kk nik gjk þ eij ; ð8Þ GE only. Hence, from Eq. (5), the GGE model is
k¼1
Yij  l  bj ¼ ai þ /ij þ eij :
the SREG is
Thus, the GGE method mixes G and GE and partitions this
X
t
Yij ¼ bj þ kk nik gjk þ eij ; ð9Þ mixture into t multiplicative terms by
k¼1 Xt
Yij  l  bj ¼ kk nik gjk :
and the AMMI model is k¼1
X
t
Note that the GGE biplot is based on the linear-bilinear
Yij ¼ l þ ai þ bj þ kk nik gjk þ eij ; ð10Þ
k¼1 SREG; see Yan and Kang (2003). As mentioned, a biplot is
as a scatterplot that graphically displays a point or score for
where t is the number of SVD axes retained in the model, each genotype and each environment. To distinguish the
kk is the singular value for the SVD axis k, nik is the sin- members of a model family, the number of components
gular value of the genotype i for the SVD axis k, gjk is the must be considered; for example, a GGE model having two
singular value of the environment j for the SVD axis k, and principal components (GGE2) and a AMMI model having
iid
eij is the error term of the models, where eij  Nð0; r2 Þ: two principal components (AMMI2). Two kinds of biplots
In all of the models in Eqs. (6), (7), (8), (9) and (10), the are common in the yield trial literature: (i) a GGE2 biplot
scale parameters kk are ordered as k1  k2      kt  0; has the first component (PC1) for its abscissa and the
and the nik and gjk satisfy the normalization and orthogo- second component (PC2) for its ordinate; and (ii) an
P 2 P 2
nality constraints, given by inik = jgjk = 1 and AMMI2 biplot has also PC1 for its abscissa and PC2 for its
P P
n g
i ik im = n g
j jk jm = 0, respectively, for k = m. ordinate. The interpretation of GGE2 and AMMI2 biplots
The COMM was used by Fisher and Mackenzie (1923), is similar. Genotypes that are more similar are closer in the
whereas the SHMM was the first linear-bilinear model used plot than genotypes that are less similar. The same is true
for identifying subsets of genotypes or environments; see for environments. Genotypes/environments that are alike

123

tend to cluster together. The angle between environmental
axes is related to the correlation between environments. An
acute angle indicates positive correlation, whereas a right
angle indicates no correlation, but an obtuse angle indicates
negative correlation. The projection of a genotype onto an
environmental axis reflects the performance of that geno-
type in that environment.
3 The GGEBiplotGUI package
In this section, we describe the R functions (commands) of
the GGEBiplotGUI package. In this package, three-
dimensional biplots are incorporated via the rgl package.
We recall that the user requires almost no knowledge of the
R programming language for using our package. The
mentioned commands are organized under the following
menu entries:
File ) View ) Biplot tools ) Format ) Models
) Biplot:
Below, we illustrate these entries, describing only those
that are not directly visible.
After the R program has been downloaded from http://
www.r-project.org and installed, it is also necessary to
install the GGEBiplotGUI package, which depends on
the rgl, tcltk and tkrplot packages, which are
automatically installed. Then, to load the GGEBiplotGUI
package into the R software, the command
library(‘‘GGEBiplotGUI’’)
must be entered at the R prompt or at any editor program
that the user is considering. Once all these instructions are
ready, the data, say, for example, data1, must be loaded;
see Sect. 4. Hence, one produces the corresponding biplot
by the command
GGEBiplot(data1)
An emerging windows, entitled ‘‘Model selection’’,
appears, from which one must select the model by choosing
the type of SVD among:
1. ‘‘JK’’ (row metric preserving);
2. ‘‘GH’’ (column metric preserving);
3. ‘‘HJ’’ (dual metric preserving); and
4. ‘‘SQ’’ (symmetrical);
the type of centering within:
1. ‘‘0’’ (no centering);
2. ‘‘1’’ (global centering: E ? G ? GE);
3. ‘‘2’’ (tester centering: G ? GE); and
4. ‘‘3’’ (double centering: GE);
and, finally, the type of scaling between:
1. ‘‘0’’ (no scaling); and
2. ‘‘1’’ (std deviation).
123
Stoch Environ Res Risk Assess
After selecting the model, one must click on the OK button
and thus the GGE biplot of data1 will be visualized by an
interactive window; for more details about these com-
mands, see Sect. 4 for an example with real data. In this
window, a menu with the following options are available:
File
• Open log file.
• Copy image: it copies the current image to the clipboard.
• Save image: it saves the current image in pdf,
postscript, metafile, bmp, png, jpg/jpeg or eps/ps
(useful for processing lateX files) formats.
• Print image: it prints the current image to any printer
connected to the user’s computer.
• Exit: it exits the program.
View
• Show both: it shows both entries and testers.
• Show genotypes: it shows genotypes only.
• Show environments: it shows environments only.
• Show/hide title.
• Show/hide guidelines.
• Add/remove symbols: it adds a dot at the location
where genotypes are placed.
Biplot tools
• Examine a genotype: it allows the user to select a
genotype. When a genotype is selected and the OK
button is clicked, the performance of different envi-
ronments in this genotype is displayed.
• Examine an environment: it allows the user to select an
environment. When an environment is selected and the
OK button is clicked, the performance of different
genotypes in this environment is displayed.
• Relation among environments: it allows us to analyze
variability and similarity/dissimilarity among environ-
ments in relation to the yield provided by the genotypes.
• Compare two genotypes: it allows a visual comparison of
two genotypes with regard to each of the environments.
When this function is chosen, a selection panel appears,
which allows one to select two genotypes to be compared.
• Which won where/what: it allows an identification of
the best genotypes in each environment.
• Discriminativeness against representativeness: it
defines an average environment and draws an aver-
age-environment-axis (AEA). It allows the representa-
tiveness and discriminating power of environments to
be visualized.
• Mean against stability: it defines an average environ-
ment and draws an AEA. It facilitates visualization of
the mean performance and stability of a cultivar.
• Rank environments with ref. to the ideal environment:
it defines an ideal environment and compares all
Stoch Environ Res Risk Assess
environments to it. The ideal environment is defined as
the most discriminating and absolutely representative.
It generates a ranking of the test environments in terms
of both criteria.
• Rank genotypes with ref. to the ideal genotype: it defines
an ideal genotype and compares all genotypes to it. The
ideal genotype is defined as one that has the highest
performance in all environments and is, therefore,
absolutely stable. It generates a ranking of the cultivars
in terms of both mean performance and stability.
• Back to original data: it resets the biplot based on the
original parameters.
Format
• Biplot title: it allows the biplot title to be modified.
• Change color with the options: background, genotypes
labels, environment labels, and biplot title.
• Change font with the options: default, larger, and smaller.
Models
This option allows us to modify the entries initially
chosen about ‘‘Model selection’’, which we now detail:
• Scaled (divided by):
– no scaling;
– std deviation;
• Centered by:
– no centering (it corresponds to the model ‘‘COMM’’);
– global-centered E ? G ? GE (it corresponds to the
model ‘‘SHMM’’);
– tester-centered G ? GE (it corresponds to the model
‘‘GGE’’);
– double-centered GE (it corresponds to the model
‘‘AMMI’’);
• SVP (singular value partitioning):
– JK (the singular value partitioning method is ‘‘row
metric preserving’’);
– GH (the singular value partitioning method is
‘‘column metric preserving’’);
– HJ (the singular value partitioning method is ‘‘dual
metric preserving’’);
– SQ (the singular value partitioning method is
‘‘symmetrical’’);
Biplot
• PC1 against PC2 (default): it shows the first principal
component (PC1) against the second principal compo-
nent (PC2).
• PC3 against PC4: it shows the third principal compo-
nent (PC3) against the fourth principal component
(PC4).
• PC5 against PC6: it shows the fifth principal compo-
nent (PC5) against the sixth principal component
(PC6).
• PC1 against PC3: it shows the first principal component
(PC1) against the third principal component (PC3).
• PC2 against PC3: it shows the second principal
component (PC2) against the third principal component
(PC3).
• Biplot3D: it shows the three dimensional biplot.
4 Data analysis
In this section, we illustrate the GGEBiplotGUI package
by using real agricultural data. These data correspond to
the yield from the 1993 Ontario winter wheat (Triticum
aestivum L.) performance trials, in which 18 cultivars were
tested at nine locations; see Yan and Kang (2003). The
data, that we name ‘ontario’ data (or simply ontario),
are displayed in Table 1.
Data may be entered using R commander or be
imported into R and saved as a matrix or a data frame.
Ontario data have been included in the package as a data
frame.
To initialize the GUI with ontario data, the
command:
GGEBiplot(Data = Ontario)
must be entered, which opens the ‘‘Model selection’’
window; see Fig. 1.
After selecting the model considering the options
‘‘GH’’, ‘‘tester-centered G ? GE’’ and ‘‘no scaling’’, one
clicks on the OK button and the main window appears;
see Fig. 2(left). The GGE biplot contains markers for
each of the 18 cultivars in lower case letters, as distin-
guished from markers for each of the nine environments
in upper case letters. The percentages of GGE explained
by the two axes are indicated in Fig. 2(left). The
GGEBiplotGUI package provides options to view the
GGE biplot in numerous ways, in order to address most
questions a breeder or researcher is likely to ask. This is
exemplified below.
4.1 The performance of different cultivars in a given
environment
From the Biplot tools menu bar, select ‘‘Examine an
environment’’. Choose any environment of interest (BH93,
in this example) from the list in the combo-box, and then
the following features will appear (see Fig. 2, right):
• A line that passes through the plot origin and the
selected environment (BH93, in this example), which is
referred to as the environment axis.
123
 Stoch Environ Res Risk Assess

Table 1 Mean yield of 18 winter wheat varieties (genotype) tested at nine Ontario locations (environment) for ontario data
Genotype Environment
BH93 EA93 HW93 ID93 KE93 NN93 OA93 RN93 WP93

Ann 4.46 4.15 2.85 3.08 5.94 4.45 4.35 4.04 2.67
Ari 4.42 4.77 2.91 3.51 5.70 5.15 4.96 4.39 2.94
Aug 4.67 4.58 3.10 3.46 6.07 5.03 4.73 3.90 2.62
Cas 4.73 4.75 3.38 3.90 6.22 5.34 4.23 4.89 3.45
Del 4.39 4.60 3.51 3.85 5.77 5.42 5.15 4.10 2.83
Dia 5.18 4.48 2.99 3.77 6.58 5.05 3.99 4.27 2.78
Ena 3.38 4.18 2.74 3.16 5.34 4.27 4.16 4.06 2.03
Fun 4.85 4.66 4.43 3.95 5.54 5.83 4.17 5.06 3.57
Ham 5.04 4.74 3.51 3.44 5.96 4.86 4.98 4.51 2.86
Har 5.20 4.66 3.60 3.76 5.94 5.35 3.90 4.45 3.30
Kar 4.29 4.53 2.76 3.42 6.14 5.25 4.86 4.14 3.15
Kat 3.15 3.04 2.39 2.35 4.23 4.26 3.38 4.07 2.10
Luc 4.10 3.88 2.30 3.72 4.56 5.15 2.60 4.96 2.89
M12 3.34 3.85 2.42 2.78 4.63 5.09 3.28 3.92 2.56
Reb 4.38 4.70 3.66 3.59 6.19 5.14 3.93 4.21 2.93
Ron 4.94 4.70 2.95 3.90 6.06 5.33 4.30 4.30 3.03
Rub 3.79 4.97 3.38 3.35 4.77 5.30 4.32 4.86 3.38
Zav 4.24 4.65 3.61 3.91 6.64 4.83 5.01 4.36 3.11

average from those performing above average in BH93.

Cultivars ‘kat’, ‘m12’, ‘ena’, ‘luc’, and ‘ann’ performed
below average, whereas other cultivars, on the same side of
the perpendicular line as BH93, performed above average.

4.2 The relative adaptation of a given cultivar

in different environments

From the Biplot tools menu bar, select ‘‘Examine a

Fig. 1 Window to select the model
• A line that passes through the plot origin and is
perpendicular to the environment axis, referred to as the
perpendicular line.
• The projections of the cultivars to the environment axis.
Figure 2(right) is based on a ‘‘Tester-centered (G ?
GE)’’ table, without any scaling and it is row metric pre-
serving (JK biplot). The cultivars are ranked in the direc-
tion of the environment axis. In this example cultivar ‘fun’
was the best, followed by ‘cas’ and ‘har’, whereas ‘kat’
was the poorest in the selected environment BH93. The
perpendicular line separates cultivars that performed below
genotype’’. Choose any genotype of interest (fun, in this
example) from the list in the combo-box and then the
following features will appear (see Fig. 3, left):
• A line that passes through the plot origin and the select
cultivar, which is referred to as the genotype axis.
• A line that passes through the plot origin and is
perpendicular to the genotype axis, referred to as the
perpendicular line.
• The projections of the environments to the genotype axis.
Figure 3(left) is based on a ‘‘Tester-centered (G ? GE)’’
table, without any scaling and it is row metric preserving.
The environments are ranked in the direction of the
genotype axis in terms of the relative performance of ‘fun’.
Thus, ‘fun’ performed the best in WP93, followed by
NN93, BH93, RN93, ID93, HW93, EA93, KE93 and
OA93. The perpendicular line separates environments, in
which ‘fun’ performed above average from those in which
‘fun’ performed below average.

123
Stoch Environ Res Risk Assess

Fig. 2 Main window for ontario data (left) and ranking cultivars based on performance in the environment BH93 for ontario data (right)

Fig. 3 Ranking environments based on relative performance of a cultivar—fun—for ontario data (left) and comparing ‘fun’ and ‘zav’ cultivars
for ontario data (right)

4.3 Comparison of two cultivars
From the Biplot tools menu bar, select ‘‘Compare two
genotypes’’. Choose any cultivar from Genotype 1 (fun),
then choose a different cultivar from Genotype 2 (zav), and
click the OK button. Upon this clicking, the following
features will appear (see Fig. 3, right):
• Two ovals that circle the two selected cultivars.
• A line that connects the two cultivars (jointer line).
• A line that is perpendicular to the jointer line and
passes through the plot origin (equality line).
Figure 3(right) is based on a ‘‘Tester-centered (G ?
GE)’’ table, without any scaling and it is row metric pre-
serving. A cultivar has higher values in environments that
are located at its side of the equality line; see Yan and
Tinker (2006). In Fig. 3(right), cultivars ‘fun’ and ‘zav’
were compared. We see that the environments OA93 and
KE93 were on the ‘zav’ side of the equality line. Thus,

123

‘zav’ was better in these environments, but ‘fun’ was better
than ‘zav’ in the other seven environments.
4.4 Which-won-where
From the Biplot tools menu bar, select ‘‘Which won where/
what’’. The GGE biplot becomes like Fig. 4(left), which is
based on a ‘‘Tester-centered (G ? GE)’’ table, without any
scaling and it is row metric preserving. The polygon is formed
by connecting the markers of the genotypes that are farthest
away from the biplot origin, such that all other genotypes are
contained in the polygon. Figure 4(left) also contains a set of
lines perpendicular to each side of the polygon. These per-
pendicular lines divide the biplot into several sectors. The
winning genotype for each sector is the one located at the
respective vertex. Genotypes located at the vertices of the
polygon reveal the best or the poorest in one or other envi-
ronment; see Yan and Tinker (2006). There are five sectors
with cultivars ‘fun’, ‘zav’, ‘ena’, ‘kat’ and ‘luc’ as the corner
or vertex cultivars. Environments OA93 and KE93 fell in the
sector in which ‘zav’ was the vertex cultivar. This means that
‘zav’ was the best cultivar for OA93 and KE93. The other
seven environments fell in the sector in which ‘fun’ was the
vertex cultivar, meaning that ’fun’ was the best cultivar for
these seven environments. No environments fell into sectors
with ‘luc’, ‘ena’, and ‘kat’ as the vertices, indicating that these
cultivars were not the best in any of the environments.
4.5 Mean performance and stability of the genotypes
From the Biplot tools menu bar, select ‘‘Mean against
stability’’. Figure 4(right) is based on a ‘‘Tester-centered
Fig. 4 The which-won-where view of the GGE biplot for ontario data (left) and AEC view for ontario data (right)
123
Stoch Environ Res Risk Assess
(G ? GE)’’ table, without any scaling and it is row metric
preserving. This figure is the average-environment coor-
dination (AEC) view of the GGE biplot, which has the
following interpretation (see Yan and Tinker 2006):
• The single-arrowed line is the AEC abscissa (or AEA)
and points to higher mean yield across environments.
Thus, ‘fun’ had the highest mean yield, followed by ‘cas’,
‘har’, etc., whereas ‘kat’ had the lowest mean yield.
• The AEC ordinate passes the plot origin and is perpen-
dicular to the AEC abscissa and points to greater
variability (poorer stability) in either direction. Thus,
‘rub’ was highly unstable, whereas ‘cas’ was highly stable.
4.6 Ranking genotypes relative to the ideal genotype
From the Biplot tools menu bar, select ‘‘Rank genotypes
with ref. to the ideal genotype’’. The arrow is where an
ideal cultivar should be. Its projection on the AEA was
designed to be equal to the longest vector of all cultivars,
and its projection on the AEC was obviously zero, meaning
that it is absolutely stable. Therefore, genotypes located
closer to the ideal genotype are more desirable than others.
Thus, ‘fun’ was more desirable than ‘cas’, where this last
was, of course, the poorest genotype; see Fig. 5(left). This
figure is based on a ‘‘Tester-centered (G ? GE)’’ table,
without any scaling and it is row metric preserving.
4.7 The representativeness and discriminating ability
of the environments
Figure 5(right) shows the representativeness and discrimi-
nating ability of environments and is based on a ‘‘Tester-
Stoch Environ Res Risk Assess

Fig. 5 Ranking cultivars based on both mean performance and stability for ontario data (left) and discriminativeness against representativeness
of test environments for ontario data (right)

centered (G ? GE)’’ table, without any scaling and it is

column metric preserving (GH biplot). The vector length,
that is, the absolute distance between the marker of an
environment and the plot origin, is a measure of the dis-
criminating ability: as the longer vector, the discrimination
of the environment increases. Therefore, among the nine
environments studied, KE93 and OA93 were most dis-
criminating (informative) and RN93 least discriminating.
The average environment (represented by the small circle
at the end of the arrow) has the average coordinates of all
test environments. AEA is the line that passes through the
average environment and the biplot origin. A test envi-
ronment that has a smaller angle with the AEA is more
representative of other test environments; see Yan and
Tinker (2006). Thus, BH93 is most representative, whereas
OA93 is least representative.

4.8 Environment ranking

The ideal test environment should be most discriminating
(informative) and also most representative of the target
environment. Figure 6 defines an ideal test environment,
which is the center of the concentric circles. This is a point
on the AEA in the positive direction (most representative),
with a distance to the biplot origin equal to the longest
vector of all environments (most informative); see Yan and
Tinker (2006). BH93 is closest to this point and is, there-
fore, best, whereas KE93 and OA93 were poorest for
selecting cultivars adapted to the whole region. Figure 6 is
based on a ‘‘Tester-centered (G ? GE)’’ table, without any
scaling and it is column metric preserving.
Fig. 6 Ranking environments based on discriminating ability and
representativeness for ontario data
5 Conclusions
In this paper, we have proposed a new interactive R
package called GGEBiplotGUI, which has implemented
the main functions of the biplot methods. This package is a
tool for multi-environment trial data visual analysis and
offers a free alternative to the existing commercial soft-
ware. In addition, the package proposed in this paper is the
only open source and interactive software on biplot

123

methods, with a graphical interface that allows us to use it
without necessity of having a knowledge of the R pro-
gramming language, which can do this package very
attractive for diverse practitioners. The GGEBiplotGUI
package graphically addresses the main questions that a
researcher need likely to ask. Although the biplot method
for genotype main effect and genotype by environment
interaction was developed originally for analyzing data of
multi-environment trials, it can also be used to visualize
other types of two-way data. For example, it was satis-
factorily employed to visualize diallel cross data (see Yan
and Hunt 2002), genotype 9 trait data (see Yan and Rajcan
2002), and genotype 9 genetic marker data (see Yan and
Falk 2002). The biplot method for genotype main effect
and genotype by environment interaction is equally appli-
cable to all types of two-way data that assume an
entry 9 tester structure. Thus, the GGEBiplotGUI
package can provide a response to several types of prob-
lems, but it mainly is intended to the analysis of data
generated by plant breeders and geneticists, in order to
visually study yields from genotypes and interactions
between genotype and environment. Another interesting
feature of this package of free distribution is that it allows
us to produce the classical biplots, which can be other
attractive aspect for different practitioners.
Acknowledgments The authors wish to thank the Editor-in-Chief,
Professor George Christakos, an Associate Editor, and anonymous
referees for their comments on an earlier version of this manuscript,
which resulted in this improved version. The research of Victor Leiva
was partially supported by FONDECYT 1120879 grant from the
Chilean government.
References
Bacon-Shone J (2008) Compositional data analysis in the geosci-
ences: from theory to practice (by A. Buccianti, G. Mateu-
Figueras and V. Pawlowsky-Glahn, eds). Stoch Environ Res
Risk Assess 22:139–141
Barros M, Paula GA, Leiva V (2009) An R implementation for
generalized Birnbaum–Saunders distributions. Comput Stat Data
Anal 53:1511–1528
Bradu D, Gabriel KR (1978) The biplot as a diagnostic tool for
models of two-way tables. Technometrics 20:47–68
Cárdenas O, Galindo MP (2003) Biplot with external information
based on generalized bilinear models. Printed by Council of
Scientific and Humanistic Development of the Central Univer-
sity of Venezuela, Caracas. Spanish version can be downloaded
from bit.ly/14BARON
Choulakian V (1966) Generalized bilinear models. Psychometrika
61:271–283
Cornelius PL, Seyedsadr MS, Crossa J (1992) Using the shifted
multiplicative model to search for separability in crop cultivar
trials. Theor Appl Genet 84:161–172
Cornelius PL, Crossa J, Seyedsadr MS (1996) Statistical tests and
estimators for multiplicative models for genotype-by-environ-
ment interaction. In: Kang MS, Gauch HG Jr (eds) Genotype-by-
environment interaction. CRC Press, Boca Raton
123
Stoch Environ Res Risk Assess
Crossa J, Cornelius JL (1997) Sites regression and shifted multipli-
cative model clustering of cultivar trial sites under heterogeneity
of error variances. Crop Sci 37:405–415
Denis JB (1991) Ajustements de modelles lineaires et bilineaires sous
constraintes lineaires avec donnes manquantes. Stat Appl
39:5–24
Ebdon JS, Gauch HG (2002a) Additive main effect and multiplicative
interaction analysis of national turfgrass performance trials I:
interpretation of genotype 9 environment interaction. Crop Sci
42:489–496
Ebdon JS, Gauch HG (2002b) Additive main effect and multiplicative
interaction analysis of national turfgrass performance trials II:
cultivar recommendations. Crop Sci 42:497–506
Eberhart SA, Russell WA (1969) Yield stability for a 10-line diallel of
single-cross and double-cross maize hybrids. Crop Sci 9:357–
361
Eckart C, Young G (1939) A principal axis transformation for non-
Hermitian matrices. Bull Am Math Soc 45:118–121
Falguerolles A (1995) Generalized bilinear models and generalized
biplots: some examples. Publications du Laboratoire de Statis-
tique et Probabilites. Universite Paul Sabatier, Toulouse
Finlay KW, Wilkinson GN (1963) The analysis of adaptation in a
plant-breeding programme. Aust J Agric Res 14:742–754
Fisher RA, Mackenzie WA (1923) The manurial response of different
potato varieties. J Agric Sci 23:311–320
Gabriel KR (1971) The biplot graphic display of matrices with application
to principal component analysis. Biometrika 58:453–467
Gabriel KR (1998) Generalised bilinear regression. Biometrika
85:689–700
Galindo MP (1986) An alternative for simultaneous representation:
HJ-biplot. Questı́io 10:12–23
Gauch HG (1988) Model selection and validation for yield trials with
interaction. Biometrics 44:705–715
Gauch HG (2006) Statistical analysis of yield trials by AMMI and
GGE. Crop Sci 46:1488–1500
Gauch HG, Zobel RW (1996) AMMI analysis of yield trials. In:
Gauch HG, Kang MS (eds) Genotype-by-environment interac-
tion. CRC Press, Boca Raton, pp 1–40
Gauch GH, Zobel RW (1997) Interpreting mega-environments and
targeting genotypes. Crop Sci 37:311–326
Gauch HG, Piepho HP, Annicchiarico P (2008) Statistical analysis of
yield trials by AMMI and GGE: further considerations. Crop Sci
48:866–889
Golub GH, Reinsch CH (1970) Singular value decomposition and
least squares solution. Numer Math 14:403–420
Gower JC (1992) Generalized biplots. Biometrika 79:475–493
Gower JC, Hand D (1996) Biplots. Chapman & Hall/CRC, London
Gower JC, Harding SA (1988) Nonlinear biplots. Biometrika
75:445–455
Gower JC, Gardner-Lubbe S, Le Roux N (2011) Understanding
biplots. Wiley, New York
Greenacre M (2010) Biplots in practice. BBVA Foundation, Madrid
Householder AS, Young G (1938) Matrix approximation and latent
roots. Am Math Mon 45:165–171
Kang MS (1988) Using genotype-by-environment interaction for crop
cultivar development. Adv Agron 62:199–252
Kempton RA (1984) The use of biplots in interpreting variety by
environment interactions. J Agric Sci 103:123–135
Leiva V, Hernandez H, Sanhueza A (2008) An R package for a
general class of inverse Gaussian distributions. J Stat Softw
26:1–21
Nelder JA, Wedderburn RWM (1972) Generalized linear models. J R
Stat Soc A 135:370–384
R Team (2013) R: a language and environment for statistical
computing. R Foundation for Statistical Computing, Vienna.
Available at http://www.R-project.org
Stoch Environ Res Risk Assess
Samonte SOPB, Wilson LT, McClung AM, Medley JC (2005)
Targeting cultivars onto rice growing environments using AMMI
and SREG GGE biplot analysis. Crop Sci 45:2414–2424
Ukkelberg A, Borgen O (1993) Outlier detection by robust alternating
regressions. Anal Chim Acta 277:489–494
Van Eeuwijk F (1995) Multiplicative interaction in generalized linear
models. Biometrics 51:1017–1032
Vicente-Villardón JL, Galindo MP, Blázquez A (2006) Logistic
biplots. In: Grenacre M, Blasius J (eds) Multiple correspondence
analysis and related methods. Chapman & Hall, New York
Yan W, Falk DE (2002) Biplot analysis of host by pathogen
interaction. Plant Dis 86:1396–1401
Yan W, Hunt LA (2002) Biplot analysis of diallel data. Crop Sci
42:21–30
Yan W, Kang MS (2003) GGE biplot analysis: a graphical tool for
breeders, geneticists, and agronomists. CRC Press, Boca Raton
Yan W, Kang MS (2006) GGEbiplot. Available at http://www.
ggebiplot.com
Yan W, Rajcan I (2002) Biplot evaluation of test sites and trait
relations of soybean in Ontario. Crop Sci 42:11–20
Yan W, Tinker NA (2006) Biplot analysis of multi-environment trial
data: principles and applications. Can J Plant Sci 86:623–645
Yan W, Hunt LA, Sheng Q, Szlavnics Z (2000) Cultivar evaluation
and mega-environment investigation based on GGE biplot. Crop
Sci 40:597–605
Yan W, Kang MS, Ma B, Woods S, Cornelius PL (2007) GGE biplot
vs. AMMI analysis of genotype-by-environment data. Crop Sci
47:643–655
123